JPH0451100A - Voice information compressing device - Google Patents

Abstract

PURPOSE:To compress information on a voice efficiently by selecting a code book to be used among plural residue signal code books according to an index value obtained from an analyzed prediction coefficient through vector quantization. CONSTITUTION:An analyzing means 12 performs the linear predictive analysis of an input voice signal and outputs the analytic result and a vector quantizing means 13 outputs the index value of the best code word from a 1st code storage means 16 stored with plural linear prediction coefficient vectors. A selecting means 14 selects a code storage means to be used among plural 2nd code storage means 17 stored with plural residue signal vectors according to the index value of the code word outputted from the vector quantizing means 13. Then a synthesizing filter means 15 inputs all code words from the selected code storage means 17 and outputs a composite signal. Consequently, the information on the voice can be compressed efficiently.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to an audio information compression device that compresses information on an audio signal and transmits or stores the compressed information.

[Prior Art] In general, audio signals have a fairly high correlation between adjacent samples, and means for efficiently compressing information in audio signals using a method of spectrum prediction based on signal sequences of adjacent samples is widely used. Furthermore, voice signals have repetitive redundancy due to pitch, and means for further increasing the efficiency of information compression by pitch prediction are widely used.

A low bit rate audio encoding method using these two linear prediction methods is CELP (Code-Excited Linear Prediction).
ed Linexr Prediction)) is being actively researched and developed (for example, ``stochastic
Coding of Speech Signals at Valuation - Low Bit Rates: The Importance of Speech Perception (Stochasji)
c Coding of 5peech Signal
s N Vet7 L.

v Bit Rates: The Important
ce ol 5peech Pe+ception)
J, M.R. Schroeder and Binisφattal (M, R, 5hotede+ and B, S
, ^ja l) Speech-Communication (Spe
ech Communication), 4.19
85. Norjh-Holland
), pp. 155-162).

In the above-mentioned CELP, the prediction residual signal is handled as a vector having a length of 40 samples, and the prediction residual signal is encoded by compressing it at a very low bit rate using vector quantization technology.

FIG. 5 shows a CEL as an example of a conventional audio information compression device.
FIG. 6 shows a decoder corresponding to the encoder of FIG. 5.

The encoder shown in FIG. 5 will be explained below.

In the encoder shown in FIG. 5, an AD converter 51 converts an analog audio signal waveform input from an input terminal into digital data, and converts it into a vector S having a certain sample length (for example, 5m5ec) to an analyzer 52 and The signals are output to the subtracter 58, respectively.

The analyzer 52 performs linear predictive analysis on the vector S output from the AD converter 51. As a result of the analysis, the prediction coefficient vector α corresponding to the input vector S and the average amplitude g of the residual signal are calculated, and the prediction coefficient vector α
is output to the linear predictor 53 and the average amplitude g of the residual signal is output to the multiplier 54, and the prediction coefficient vector α and the average amplitude g of the residual signal are also output to the multiplexer 55. Here, linear prediction collectively refers to spectrum prediction based on signal sequences of adjacent samples and pitch prediction, which is generally referred to as long-term prediction.

The codebook 56 includes a gain-normalized residual signal waveform (hereinafter referred to as a codeword) with a constant sample length.
A plurality of types (for example, 1024 types) are stored, and the stored code words are output to the multiplier 54 according to instructions from the optimal code word selector 57.

The multiplier 54 amplifies the code word output from the codebook 56 by the average amplitude g of the residual signal output from the analyzer 52, and converts it into a synthesis filter composed of a linear predictor 53, an adder 61, etc. Output to.

The synthesis filter receives the amplified codeword and outputs a synthesized signal vector S'.

The subtracter 58 inputs this composite signal vector S' and the input vector S and calculates a difference.

The auditory weighting filter 59 considers the effect of auditory spectrum masking on the difference output from the subtracter 58 and outputs it to the power calculator 60 .
The power is calculated and sent to the optimal codeword selector 57. The optimal codeword selector 57 selects the codeword with the minimum power from the codebook 56 and sets an index value 1nd for specifying the codeword.
output ex to multiplexer 55.

The multiplexer 55 transmits the prediction coefficient vector α, the average amplitude g of the residual signal, and the index value 1ndex for specifying the residual signal.

Here, methods for creating the codebook 56 that stores codewords include a method in which an actual audio signal is analyzed for a long time to obtain a residual signal, and a method in which the residual signal is created by vector quantization, or a method in which the codebook 56 is created using a White-Gaussian method. (White-Ga
There is a method of creating a pseudo signal using a random signal of ``ussian''.

Next, the decoder shown in FIG. 6 will be explained.

The demultiplexer 62 decomposes the signal received from the transmission path into a prediction coefficient vector α, an average amplitude g of the residual signal, and an index value 1ndex for specifying the residual signal.

The codebook 63 has the same contents as the codebook 56 of the encoder, and a codeword is output to the multiplier 64 based on the index value 1ndex that specifies the residual signal.

The multiplier 64 converts the codeword received from the codebook 63 into an average amplitude g output from the demultiplexer 62.
Amplify by.

A synthesis filter including an adder 65, a linear predictor 66, etc. calculates a synthesis signal vector using a synthesis filter formed according to the received prediction vector α.

The DA converter 67 inputs the composite signal vector output from the synthesis filter, converts the reproduced signal into an analog signal, and reproduces the input audio signal.

[Problems to be Solved by the Invention] However, in the above-mentioned speech encoding method using CELP, the content of the codebook is only one type, and the codebook is set to be optimized based on the results of long-term statistical analysis of speech. has been done. Therefore, it is known that the codobook resulting from long-term statistical analysis is not optimal at the point of plosive consonant plosive consonants, silence, and the transition period from voicelessness to vowel stationary, etc. In the general CELP that is expressed, the residual signal of the nasal interval contains information on the zero points of the spectrum that could not be expressed by the all-pole model, and the code that is the result of long-term statistical analysis. There is a problem that books cannot adequately address.

In view of the above-mentioned problems of the conventional speech information compression device, it is an object of the present invention to provide a plurality of residual signal codebooks suitable for various short-time speech patterns, and to provide an index that vector quantizes analyzed prediction coefficients. An object of the present invention is to provide an audio information compression device that can efficiently compress audio information by selecting a codebook to be used from among a plurality of residual signal codebooks based on the values.

[Means for Solving the Problem] The above-mentioned object of the present invention is to provide an analysis means for performing linear predictive analysis of an input audio signal and outputting the analysis result, and a first code for storing a plurality of linear predictive coefficient vectors. a storage means, a vector quantization means for obtaining an optimum codeword index value from the first code storage means, a plurality of second code storage means each storing a plurality of residual signal vectors, and a vector quantization means from the vector quantization means. a selection means for selecting a code storage means to be used from a plurality of second code storage means based on the output index value of the optimum code word; and inputting all code words from the code storage means selected by the selection means. This is achieved by an audio information compression apparatus comprising a synthesis filter means for outputting a synthesis signal.

[Operation] The analysis means performs linear prediction analysis on the input audio signal and outputs the analysis result, and the vector quantization means extracts an optimal code word from the first code storage means storing a plurality of linear prediction coefficient vectors. the selection means outputs an index value of the optimal codeword outputted from the vector quantization means, and uses the selection means from a plurality of second code storage means each storing a plurality of residual signal vectors based on the index value of the optimal codeword outputted from the vector quantization means. The synthesis filter means inputs all the code words from the code storage means selected by the selection means and outputs a synthesized signal.

[Example] Hereinafter, an example of the audio information compression apparatus according to the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of an encoder as a first embodiment of the audio information compression apparatus of the present invention, and FIG. 2 is a block diagram showing a decoder corresponding to the encoder shown in FIG. FIG. 2 is a block diagram showing the configuration of FIG.

The encoder shown in FIG. 1 includes an AD converter 11. an analyzer 12 as an analysis means connected to the AD converter 11; a prediction coefficient vector quantizer (represented as prediction coefficient VQ in the figure) I3 as a vector quantization means connected to the analyzer 12;
A codebook selector 14 as a selection means connected to the prediction coefficient vector quantizer 13, a linear predictor 15 connected to the prediction coefficient vector quantizer I3 and the codebook selector 14, and a prediction coefficient vector quantizer 13. and a prediction coefficient codebook 16 as a first code storage means connected to the linear predictor 15, a residual signal codebook 17 as a second code storage means, connected to the analyzer 12 and the residual signal codebook 17. A multiplier 18, an adder 19 connected to the multiplier 18 and the linear predictor 15, and an AD converter 1
1. A subtracter 20 connected to the adder 19 and the linear predictor 15, a filter 21 for acoustic weighting connected to the subtractor 20, a power calculator 22 connected to the filter 21 for acoustic weighting, and a power calculator 22 connected to the filter 21 for acoustic weighting. It consists of an optimal codeword selector 23 connected to the calculator 22, and a multiplexer 24 connected to the analyzer 12, the prediction coefficient vector quantizer 13, and the optimal codeword selector 23.

Next, the operation of the above encoder will be explained in detail.

First, the AD converter 11 converts an analog audio signal waveform input from an input terminal into digital data, and outputs the data as a vector S with a constant sample length (for example, 5 m5 ec) to the analyzer 12 and the subtracter 20, respectively.

Analyzer I2 performs linear predictive analysis on vector S output from AD converter II. Here, linear prediction collectively refers to spectrum prediction based on signal sequences of adjacent samples and pitch prediction, which is generally referred to as long-term prediction. Then, the analyzer 12 calculates the prediction coefficient vector α and the average amplitude g of the residual signal as the analysis result, and sends the prediction coefficient vector α to the prediction coefficient vector quantizer 13 to calculate the average amplitude g of the residual signal. Each output is output to a multiplier 18.

The prediction coefficient vector quantizer 13 refers to the prediction coefficient codebook 16 to obtain an optimal codeword index value j.

The codebook selector 14 selects a codebook to be used from the plurality of residual signal codebooks 17 based on the index value j output from the prediction coefficient vector quantizer 13. Note that a method for generating the plurality of residual signal codebooks 17 and a method for selecting a codebook to be used will be described later.

The residual signal codebook I7 selected by the codebook selector 14 stores a plurality of types (for example, 256 types) of codewords in which the gain of a constant sample length is normalized.

The optimal codeword selector 23 is the codebook selector 1
For the residual signal codebook selected by 4,
Instructs multiplier 18 to output all codewords stored in the selected residual signal codebook.

After amplifying the code word output from the residual signal codebook selected by the multiplier 18, the linear predictor 1
5 and the like calculates a composite signal vector S' from the amplified need word and outputs it to the subtracter 20.

A subtracter 20 calculates an error by subtracting the composite signal vector S' from the input vector S, and outputs the auditory weighting to a filter 21. The filter 21 takes into account the effect of spectral masking on the input error and outputs the auditory weighting to the power calculator 22.
The power is calculated, and the optimal codeword selector 23
An index value i is obtained that specifies the codeword with the minimum power in the selected codebook.

The multiplexer 24 transmits to the communication path 25 the index value j1 that specifies the prediction coefficient vector, the average amplitude g of the residual signal, and the index value i that specifies the residual signal.

Next, creation of a plurality of codebooks 17 that store residual signal vectors suitable for the characteristics of the audio signal described above will be described.

In the encoder of this embodiment, the creation of codebooks is a very important issue, and if it is not possible to appropriately create multiple codebooks with multiple codewords, it will be necessary to improve the encoding performance. cannot be expected. Another disadvantage is that the amount of memory for storing multiple codebooks increases.

Hereinafter, as an example of a method for creating a plurality of codebooks 17 having a plurality of codewords and a method for selecting a food book to be used, a procedure for creating the codebooks 17 by analyzing an actual audio signal over a long period of time will be described.

Step 1: In the same way as the conventional procedure, create a large size (eg, N=4096 codewords) residual signal codebook by analyzing long-term speech.

Then, the index value of this code word is n (n=1.2.
-, N) (however, N is a positive integer).

Step 2: The prediction coefficient vector quantizer 13 uses the residual signal codebook created in Step 1 to
Encode the long speech again.

At this time, the index value of vector quantization of the linear prediction coefficient is j
(j = 1.2....9M, where M is an arbitrary positive integer representing the codebook size of the linear prediction coefficient), all codes of the residual signal are calculated for each index value j. Find contribution level for words. Here, the degree of contribution is
It shows how often a certain code word was used and what its performance (quality of synthesized speech) was at that time.

As an example, the index value j of vector quantization of linear prediction coefficients
If the contribution degree of the code word of index value n of the residual signal in is C(D, then C(Dn
n=(probability that index value n of the residual signal is used)×(average performance of synthesized speech using index value n of the residual signal).

Average performance here refers to the signal-to-noise ratio (signal-to-noise ratio).
eRatio)).

Step 3: Having a codebook of residual signals for each vector quantization index value j of all linear prediction coefficients is
This is unrealistic in that it requires a huge amount of memory. Therefore, for each index value j of vector quantization of linear prediction coefficients,
Collect codebooks with similar usage status,
It is necessary to reduce the contribution patterns of the codewords of the residual signal to a small number.

Such a summary can be summarized by considering the contribution C (D) as a vector of N samples with variable J, and using the method of vector quantization, K is a number smaller than J, and K types (where K is , the number of codebook types of the residual signal) by obtaining contribution data C(k) (k=1. 2..., K).

In other words, assuming the degree of contribution as C(k), k=1.2°...
Let K<j.

During this vector quantization, the correspondence between the vector quantization index value j of the linear prediction coefficients and k summarized by vector quantization is stored. An example of this correspondence is shown in Table 1 (here, J is 1024 and K is 8).

Through steps 1 to 3 described above, the usage status of the codebook of the residual signal is clustered for each feature of the audio signal.

Step 4: The final step is to reduce the size of the codebook of the residual signal of the type described above.

Up to step 3, the contents of the codebooks of residual signals of all clusters are the same, but the degree of contribution of codewords differs for each cluster.

Therefore, by extracting codewords with a high degree of contribution from each codebook, the size of each foodbook is reduced. Here, the codebook size of the residual signal in step 1 was 4096 types (the index value is expressed as 12bits), and the codebook size was 4096 types, and the codebook size was 4096 types, and the codebook size was 4096 types (the index value was expressed as 12bits), and 1
024 Narrow down to types (index value is expressed in 10 bits).

By the above-described procedure, an appropriate codebook of residual signals is created for each feature of the audio signal, and the created codebook can also be selected.

Next, FIG. 2 shows a decoder corresponding to the encoder of FIG. 1.

The decoder of FIG. 2 includes a demultiplexer 26, a codebook selector 27 connected to the demultiplexer 26, a residual signal codebook 28 connected to the demultiplexer 26, and a demultiplexer 26 and a codebook 28 connected to each other. a multiplier 29, an adder 3 connected to the multiplier 29;
0. A linear predictor 31 connected to the adder 3o. It is composed of a prediction coefficient codebook 32 connected to a linear predictor 31, an adder 30, and a DA converter 33 connected to the linear predictor 31.

Next, the operation of the decoder shown in FIG. 2 will be explained.

The demultiplexer 26 decomposes the signal received from the transmission path 25 into an index value j that specifies the prediction coefficient vector α, an average amplitude g of the residual signal, and an index value i that specifies the residual signal.

The codebook selector 27 selects a residual signal codebook to be used from a plurality of residual signal codebooks 28 provided in advance, based on an index value j that specifies the prediction coefficient vector α. Note that this operation is the same as that of an encoder. Furthermore, the plurality of residual signal codebooks 28 and predictive coefficient codebooks 32 have the same contents as the codebooks 17 and 16 of the encoder, respectively.

When the residual signal codebook to be used is selected by the codebook selector 27, the codeword specified by the index value i of the selected residual signal codebook is added to the multiplier 2.
Output to 9.

The multiplier 29 inputs the specified code word, amplifies the code word using the average amplitude g output from the demultiplexer 26, and outputs the code word to a synthesis filter composed of an adder 30, a linear predictor 31, etc. do.

The synthesis filter calculates a synthesis signal vector from the codeword amplified by the synthesis filter, which is formed according to the index value j of the prediction coefficient codeword output from the demultiplexer 26 , and outputs it to the DA converter 33 .

The DA converter 33 converts the input composite signal vector into an analog signal and reproduces the input audio signal.

According to the above operating procedure, the codebook of the residual signal is changed based on the characteristics of the audio signal, so that audio information can be efficiently compressed.

In the first embodiment, by providing a plurality of residual signal codebooks, the amount of memory required increases.

Furthermore, as understood from the procedure explained in the method for generating multiple residual signal foodbooks described above, each residual signal codebook is extracted from a large (for example, 4096 types) residual signal codebook (for example, 1024 types). ). Note that there are overlapping codewords in the contents of the plurality of residual signal codebooks.

Next, FIG. 3 shows a block diagram of the configuration of an encoder as a second embodiment of the audio information compression apparatus of the present invention, and FIG. 4 shows a decoder corresponding to the encoder shown in FIG. 3. A block diagram of the configuration is shown.

In view of the above points in the first embodiment, the encoder shown in FIG. 3 is configured to reduce the amount of memory required for the plurality of residual signal codebooks.

The difference between the encoder shown in FIG. 3 and the decoder shown in FIG. 4 from the encoder shown in FIG. 1 and the decoder shown in FIG. 2 is as follows.
Multiple residual signal codebooks 17 in FIGS. 1 and 2
．． 28 is replaced by a plurality of residual signal codebooks 34 and 36 having different contents, and a residual signal codebook 3537 is newly added.

Other than the above-mentioned parts, the operation is exactly the same as in the first embodiment, so these residual signal codebooks 34, 35, 36 will be described below.
．． Only 37 will be explained.

In the encoder shown in FIG. 3 and the decoder shown in FIG. 4, the only codebooks that actually store residual signals are residual signal food books 35 and 37. The plurality of residual signal codebooks 34, 36 do not directly store residual signals, but rather store residual signal codebooks 35.36. It only stores index values that identify 37 code words. Therefore, a plurality of residual signal codebooks 34. 36
The amount of memory required for this is relatively small.

Note that the code book 35. actually stores the residual signal.
The creation of the residual signal codebook 37 is performed in the creation of the plurality of residual signal codebooks of the first embodiment described above.
．． It is sufficient to restore all 36 index values and extract and create code words from the codebook of the large residual signal created in step 1 without duplication.

The encoder shown in Fig. 3 and the decoder shown in Fig. 4 efficiently provide speech information by adaptively changing the codebook of the residual signal stored in a small memory based on the characteristics of the speech signal. Can be compressed.

[Effects of the Invention] An analysis means that performs a linear prediction analysis of an input audio signal and outputs the analysis result, a first code storage means that stores a plurality of linear prediction coefficient vectors, and an optimal a vector quantization means for obtaining an index value of a codeword, a plurality of second code storage means each storing a plurality of residual signal vectors, and an index of the optimal codeword output from the vector quantization means. selection means for selecting a code storage means to be used from the plurality of second code storage means based on the value; and synthesis for inputting all code words from the code storage means selected by the selection means and outputting a composite signal. Since it is equipped with a filter means, it is possible to select and use an appropriate residual signal code storage means without using auxiliary information for selecting a code storage means, and as a result, it is possible to efficiently record audio. Information can be compressed.

Figure 2

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of an encoder as a first embodiment of the audio information compression apparatus of the present invention, and FIG. 2 is an example of the configuration of a decoder corresponding to the encoder shown in FIG. 1. FIG. 3 is a block diagram showing the configuration of an encoder as a second embodiment of the audio information compression apparatus of the present invention, and FIG. 4 is a decoding diagram corresponding to the encoder shown in FIG. 3. FIG. 5 is a block diagram showing an example of an encoder of a conventional audio information compression device, and FIG. 6 is a block diagram showing an example of a decoder corresponding to the encoder shown in FIG. FIG. 2 is a block diagram showing an example. 11...AD converter, 12...analyzer, 13...
Prediction coefficient vector quantizer, 14.27...Codebook selector, 15, 31...Linear predictor, 16.32
...Prediction coefficient codebook, 17.28.34.36
... Residual signal food book, Ill, 29... Multiplier, 19.30... Adder, 20... Subtractor, 2
1... Auditory weighting is a filter, 22... Power calculator, 23... Optimal code word selector, 24...
Multiplexer, 26... Demultiplexer, 33...
...DA converter. Figure 3 Figure 4

Claims (1)

[Claims] an analysis means for performing a linear prediction analysis of an input audio signal and outputting an analysis result; a first code storage means for storing a plurality of linear prediction coefficient vectors; and an index of an optimal code word from the first code storage means. vector quantization means for obtaining a value; a plurality of second code storage means each storing a plurality of residual signal vectors; A selection means for selecting a code storage means to be used from a plurality of second code storage means, and a synthesis filter means for inputting all the code words from the code storage means selected by the selection means and outputting a composite signal. An audio information compression device characterized by: